Taking R to New Heights for Scalability and Performance

Transcription

1 Taking R to New Heights for Scalability and Performance Mark Hornick Director, Advanced Analytics and Machine Learning blogs.oracle.com/r January 31,2017

2 Safe Harbor Statement The following is intended to outline our general product direction. It is intended for information purposes only, and may not be incorporated into any contract. It is not a commitment to deliver any material, code, or functionality, and should not be relied upon in making purchasing decisions. The development, release, and timing of any features or functionality described for Oracle s products remains at the sole discretion of Oracle. 2

3 Why statisticians data analysts data scientists use R Powerful Extensible Graphical Extensive statistics Ease of installation and use Rich ecosystem ~10K open source packages Millions of users worldwide Free R is a statistics language similar to Base SAS or SPSS statistics 3

4 Traditional R and Data Source Interaction Read/Write using built-in R capabilities read Flat Files extract / export Data Source export load RODBC / RJDBC / ROracle Deployment R script cron job Access latency Paradigm shift: R Data Access Language R Memory limitation data size, call-by-value Single threaded Ad hoc production deployment Issues for backup, recovery, security 4

5 How to take R to new heights for scalability and performance? i.e., to work on Big Data 5

6 Volume Veracity Variety Value Variability Velocity 6

7 Capabilities that take R to new heights Transparent Data Access, Analysis, and Exploration Production Deployment Scalable Machine Learning Data & Task Parallelism 7

8 Transparent data access, analysis, and exploration All Processing Light local Processing Translate R invocations Offload processing to more powerful machines and environments using data.frame proxies Avoid data movement SQL Spark Java, Python, Scala, R HiveQL Data Source Main Processing Main Processing Main Processing 8

9 Transparent data access and manipulation Maintain language features and interface Transparently translate R to language of powerful data processing engines Reference data to eliminate data movement Analyze all of your data 9

10 Proxy objects for Big Data data.frame Proxy data.frame Inherits from 10

11 Transparency Examples library(ore) ore.connect("rquser", "orcl", "localhost", "rquser", all=true) ore.ls() hist(my_table$arrdelay,breaks=100) merge (TEST_DF1, TEST_DF2, by.x="x1", by.y="x2") df <- with(ontime_s, ONTIME_S[DEST=="SFO" DEST=="BOS",1:21]) df$lrgdelay <- ifelse(df$arrdelay > 20,1,0) head(df) summary(df) # with OREdplyr in ORE select(flights, year, month, dep_delay) rename(flights, tail_num = tailnum) filter(flights, month == 1, day == 1) arrange(flights, year, month, day) mutate(flights, speed=air_time/distance) ore.frame Proxy Object 11

12 Scalable Machine Learning Offload model building to parallel software and powerful machines and environments Light local Processing Enable new technologies as they arise All Processing RDBMS Build models using powerful hardware and parallel algorithms Spark Flink Hadoop Data.Rdata Main Processing Main Processing Main Processing Main Processing OAA ORAAH, MLlib FlinkML ORAAH, Mahout 12

13 Scalable Machine Learning Maintain R machine learning interface Easy to specify formula minimal lines of code Include transformations, interaction terms, etc. Algorithm Façade Algorithm Target Predictors log(arrdelay) ~ DISTANCE + ORIGIN + DEST + as.factor(month) + as.factor(year) + as.factor(dayofmonth) + as.factor(dayofweek) + as.factor(flightnum) 13

14 Scalable Machine Learning Maintain R machine learning interface Easy to specify formula minimal lines of code Include transformations, interaction terms, etc. Bring the algorithm to the data Eliminate or minimize data movement Leverage proxy objects to reference data Slow Fast Algorithm Data 14

15 Scalable Machine Learning Maintain R machine learning interface Easy to specify formula minimal lines of code Include transformations, interaction terms, etc. Bring the algorithm to the data Eliminate or minimize data movement Leverage proxy objects to reference data Parallel, distributed algorithm implementations Oracle-proprietary parallel, distributed algorithms Leverage other open source packages and toolkits e.g., Apache Spark Mllib, Apache FlinkML Algorithm 15

16 Linear Model Performance Comparison Predict Total Revenue of a customer based on 31 numeric variables as predictors, on 184 million records using SPARC T5-8, 4TB of RAM Data in an Oracle Database table Algorithm Threads Used* Memory required** Time for Data Loading*** Time for Computation Total Relative Performance Open-Source R Linear Model (lm) 1 220Gb 1h3min 43min 1h46min 1x Oracle R Enterprise lm (ore.lm) min 42.8min 2.47X Oracle R Enterprise lm (ore.lm) min34s 1min34s 67.7X Oracle R Enterprise lm (ore.lm) s 57.97s 110X Oracle R Enterprise lm (ore.lm) s 41.69s 153X *Open-source R lm() is single threaded **Data moved into the R Session s memory, since open-source lm() requires all data to be in-memory ***How long it takes to load 40Gb of raw data into the open-source R Session s memory 16

17 Time in Seconds Time in Seconds Not all parallel implementations are the same Comparing performance with varying Spark memory footprints Benchmark on single X5-2 Node with 74 threads and 256 GB of Total RAM, Spark on CDH Input Data is 15GB Ontime airline dataset with 123mi records, predicting 8,926 total coefficients 1, LM in Spark Performance ORAAH LM Spark MLlib LM 908 1, GLM in Spark Performance ORAAH GLM Spark Mllib GLM DNF Spark Context memory in GB WNR LM formula used ARRDELAY ~ DISTANCE + ORIGIN + DEST + as.factor(month) + as.factor(year) + as.factor(dayofmonth) + as.factor(dayofweek) + as.factor(flightnum) GLM formula used CANCELLED DISTANCE + ORIGIN + DEST + as.factor(month) + as.factor(year) + as.factor(dayofmonth) + as.factor(dayofweek) + as.factor(flightnum) Spark Context memory in GB WNR

18 Data and Task Parallel Execution Easily specify parallelism and data partitioning Simplified API all-in-one Build and score with millions of models Automated management of parallel R engines Insulation from hardware details Limit resources as appropriate Startup and shutdown automatically Automated loading of data into parallel R engines Leverage CRAN packages 18

19 Data and Task Parallelism Hand-code logic to spawn R engines and partition and feed data to R engines as they become available All Processing parallel Rserve Rmpi snow Store UDF Invoke Execute user-defined R function on back-end servers in data-parallel or task-parallel manner Auto-partition and feed data while also leveraging CRAN packages Partition data by value Partition data by count invoke function with index Data.Rdata.Rdata.Rdata.Rdata Spawn & control R engines Provide function and data R Script Repository R Object Repository 19

20 Example API Supply data Specify function Use CRAN packages Store and load R objects Pass Arguments Specify parallelism Get/use results R objects structured data Images etc. library(e1071) mod <- ore.tableapply(iris_table, function(dat,datastore) { library(e1071) dat$species <- as.factor(dat$species) mod<-naivebayes(species ~., dat) ore.save(mod,name=datastore) }, datastore="nb_model-1") scorenbmodel <- function(dat, datastore) { } library(e1071) ore.load(datastore) dat$pred <- predict(mod, newdata = dat) dat IRIS_PRED IRIS_PRED$PRED <- "A" <- IRIS_TABLE[1,] res <- ore.rowapply(iris_table, scorenbmodel, datastore = "NB_Model-1", DAT <- ONTIME_S[ONTIME_S$DEST %in% parallel=4, FUN.VALUE=IRIS_PRED, rows=10) c("bos","sfo","lax","ord","atl","phx","den"),] No parallelism Data parallel by chunk Data parallel by partition modlist <- ore.groupapply( X=DAT, INDEX=DAT$DEST, parallel=3, function(dat) lm(arrdelay ~ DISTANCE + DEPDELAY, dat)) 20

21 Deployment Avoid costly recoding or translating R code Invoke R easily from non-r environments Map data structures and types naturally Seamlessly return data.frames, images, XML, JSON in local environment data structures 21

22 Deployment data.frame images R objects Application Business Logic C, C++, Java, etc. rjava API RServe client/server TCP/IP rpy2 system() cron independent Application / Dashboards Business Logic C, C++, Java SQL Allow applications and dashboard tools to use familiar, existing SQL protocols for invoking R Invoke R from SQL results as table: structured, image, XML, JSON automate parallel and concurrent execution R Scripts Data Stores RDBMS HDFS NoSQL Hive File system SQL - R Engine R Scripts Data RDBMS HDFS NoSQL Hive 22

23 Deploy R using SQL Store named R function in Script Repository from R or SQL Return values Images as PNG BLOB column data.frame content as database table XML with data.frame and image Benefits Fewer moving parts IPC data transfer speeds at backend Invoke same function from R or SQL Security Integrated backup and recovery begin sys.rqscriptdrop('randomreddots'); sys.rqscriptcreate('randomreddots', 'function(){ id <- 1:10 plot( 1:100, rnorm(100), pch = 21, bg = "red", cex = 2, main="random Red Dots" ) data.frame(id=id, val=id / 100) }'); end; select from select from ID, IMAGE table(rqeval( NULL,'PNG','RandomRedDots')); id, val table(rqeval( NULL,'select 1 id, 1 val from dual', 'RandomRedDots')); -- Return structured and image content within XML string select * from table(rqeval(null, 'XML', 'RandomRedDots')); -- In R, invoke same function by name ore.doeval(fun.name='randomreddots') 23

24 Architectural Elements: Enabling R for Big Data Leverage powerful back-ends for the heavy lifting transparently Leverage new, more powerful back-ends more easily as they appear Enable parallelism quickly and easily for big data processing Immediately leverage data scientist R scripts and results in production environments Production Deployment Transparent Data Access, Analysis, and Exploration Data & Task Parallelism Scalable Machine Learning 24

25 Oracle R Enterprise Part of Oracle Advanced Analytics option to Oracle Database Use Oracle Database as HPC environment Use in-database parallel and distributed machine learning algorithms Manage R scripts and R objects in Oracle Database Integrate R results into applications and dashboards via SQL R Client Oracle R Enterprise Oracle Database User tables In-db stats SQL Interfaces SQL*Plus, SQLDeveloper, Database Server Machine 25

26 Oracle R Enterprise Part of Oracle Advanced Analytics option to Oracle Database Transparency layer Leverage proxy objects (ore.frames) - data remains in the database Overload R functions that translate functionality to SQL Use standard R syntax to manipulate database data Parallel, distributed algorithms Scalability and performance Oracle R Enterprise Exposes in-database algorithms from ODM Additional R-based algorithms executing and database server Embedded R execution Manage and invoke R scripts in Oracle Database Data-parallel, task-parallel, and non-parallel execution Use open source CRAN packages R Client Oracle Database User tables In-db stats SQL Interfaces SQL*Plus, SQLDeveloper, Database Server Machine 26

27 OAA / Oracle R Enterprise Predictive Analytics algorithms in-database plus open source R packages for algorithms in combination with embedded R data- and task-parallel execution Classification Decision Tree Logistic Regression Naïve Bayes Support Vector Machine RandomForest Regression Linear Model Generalized Linear Model Multi-Layer Neural Networks Stepwise Linear Regression Support Vector Machine New in ORE *ODB 12.2 only Clustering Hierarchical k-means Orthogonal Partitioning Expectation Maximization* Attribute Importance Minimum Description Length Expectation Maximization* Anomaly Detection 1 Class Support Vector Machine Market Basket Analysis Apriori Association Rules Feature Extraction Nonnegative Matrix Factorization Principal Component Analysis Singular Value Decomposition Explicit Semantic Analysis* Time Series Single Exponential Smoothing Double Exponential Smoothing 27

28 Oracle R Advanced Analytics for Hadoop Using Hadoop/Hive/Spark Integration, plus R Engine and Open-Source R Packages Oracle R Advanced Analytics for Hadoop (ORAAH) on Hadoop Cluster R interface to HQL Basic Statistics, Data Prep, Joins and View creation Parallel, distributed algorithms: ORAAH Spark algorithms: Deep Neural, GLM, LM Spark MLlib algorithms: LM, GLM, LASSO, Ridge Regression, Decision Trees, Random Forests, SVM, k-means, PCA HQL Use of Open-source R packages via custom R Mappers / Reducers Oracle Database with Advanced Analytics option R Client R Analytics Oracle R Advanced Analytics for Hadoop SQL Client SQL Developer Other SQL Apps

29 Oracle R Advanced Analytics for Hadoop Predictive Analytics algorithms Classification GLM ORAAH Logistic Regression ORAAH Logistic Regression Random Forests Decision Trees Support Vector Machines Clustering Hierarchical k-means Hierarchical k-means Gaussian Mixture Models Regression MLP Neural Networks ORAAH LASSO Ridge Regression Support Vector Machines Random Forest Linear Regression Basic Statistics Correlation/Covariance Feature Extraction Non-negative Matrix Factorization Collaborative Filtering (LMF) Singular Value Decomposition Attribute Importance Principal Components Analysis Principal Components Analysis 29

30 Cloud-Based Machine Learning Oracle Advanced Analytics option including Oracle Data Mining and Oracle R Enterprise on: Oracle Exadata Cloud Service Oracle Database Cloud Service: Included in High Performance and Extreme Performance services Oracle R Advanced Analytics for Hadoop Included in the Oracle Big Data Cloud Service 30

31 Demonstration of ORE 31

32 Join us for the Oracle R Enterprise Hands-on Lab 9:00 Using R for Big Data Advanced Analytics and Machine Learning data exploration / attribute importance clustering regression OREdplyr, and more 32

33 Join us for Big Data with ORAAH 1:00 Using Machine Learning to unlock the Business Value in Big Data 33

34 Join us for new technology intro 9:50 Combining Graph and Machine Learning Technologies using R 34

35 Join us for new technology intro 10:55 Introducing Oracle Machine Learning new notebook technology from Oracle 35

36 Learn More about Oracle s Advanced Analytics R Technologies 36

37